Joule Counting Correction for Electric Vehicles Using Artificial Neural Networks

Michael David TaylorAAAI 2014 · DOI

Introduction

Estimating Battery Remaining Energy (BRE) in electric vehicles (EVs) is a complex challenge. Unlike traditional fuel tanks, battery energy levels cannot be measured directly; they must be inferred from voltage, current, and temperature sensors. The industry standard, "coulomb counting," integrates current over time but suffers from "drift"—the accumulation of small sensor errors and unmodeled physical effects like internal heat loss. This uncertainty often leads to "range anxiety" for drivers.

This research introduces a hybrid approach that combines traditional "Joule counting" (integrating power over time) with an Artificial Neural Network (ANN). By leveraging the ANN to learn and correct non-linear errors—primarily those caused by internal resistive heating—the method achieves an accuracy of 2.5% to 5%, providing drivers with a reliable estimate within 1–2 miles of actual range.

Methodology

The proposed "Joule Counting Correction" algorithm operates in a two-step iterative cycle:

  1. Initial Estimation (Joule Counting): The system integrates power (Watts = Volts × Amps) to track energy in kWh. While more direct than counting charge (Amps), this method still accumulates error over time.
  2. Neural Network Correction: An ANN with one hidden layer (25 neurons) predicts the error accumulated by the Joule counter. The model uses instantaneous and averaged sensor data (voltage, current, temperature) as inputs.
The iterative cycle of data collection during driving/charging and model adaptation upon reaching a full charge.

To remain accurate as conditions change, the system employs Adaptive Training. The network re-trains itself after every full charge cycle using data from the most recent three cycles. This allows the model to adapt to seasonal temperature shifts or gradual battery degradation without manual recalibration.

Dataset and Experimental Setup

The study utilized a unique real-world dataset from the Charge Car project at Carnegie Mellon University. The test vehicle was a converted EV equipped with a 10kWh lithium iron phosphate (LiFePO4) battery pack.

Unlike lab-based studies, this data was collected over 81 days from four different drivers, covering 710 miles and 15 full charge/discharge cycles. This "messy" real-world data—including stop-and-go traffic, hills, and auxiliary power draws (the "vampire drain" from electronics)—provided a rigorous testing ground for the ANN's predictive capabilities.

Key Results and Insights

The researchers found that error accumulates most rapidly during heavy use (high current), mirroring the physics of resistive heating ($P_{loss} = I^2R$). While simple linear models proved too erratic to handle these non-linearities, the ANN successfully "pulled" the drifting Joule-counted estimates back toward the ground truth.

Comparison between raw Joule counting and the ANN-corrected estimate over several days of driving, showing how the correction aligns the estimate with known full-charge states.

Key performance metrics included:

  • High Accuracy: The corrected model maintained an error rate between 2.5% and 5% of total capacity.
  • Drift Correction: As shown in the performance data, uncorrected Joule counting consistently overestimated energy, while the ANN correction significantly reduced both error and variance over successive training epochs.

Limitations and Future Work

While promising, the model faces potential risks of overfitting to specific driving styles. If a new driver with a different profile takes the wheel, the model may require several cycles to re-adapt. Future work aims to incorporate State of Health (SOH) estimation to account for long-term capacity loss and integrate GPS data to factor in terrain and elevation changes for even more precise range forecasting.

Summary updated

BibTeX
@inproceedings{taylor2014joule,
  author = {Michael David Taylor},
  title = {{Joule Counting Correction for Electric Vehicles Using Artificial Neural Networks}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2014},
  doi = {10.1609/aaai.v28i1.8742},
}
Figure